In October 2026 the market said out loud what it had been pricing quietly about GPU compute. CoreWeave‘s 9.75% notes due 2031 traded at a 15.78% yield, and its five-year credit default swap spread, the annual cost of insuring against non-repayment, sat near 1,038 basis points (100 basis points is a percentage point). The same balance sheet carried $35.6 billion of debt against a $103.7 billion backlog, 72% from three customers.
None of it is on a procurement checklist, yet a multi-year commitment prices it regardless. Signing extends you credit, and the rate card prices a balance sheet you never see. This guide traces six mechanisms: signal, ownership, collateral, duration, enforcement, choice. Five levers are yours: term length, commitment size, prepayment protection, portability and provider choice.
In This Series
- What a 15.78% bond yield reveals about your GPU provider — How a 9.75% coupon trading at a 15.78% yield and a 1,038 basis point spread are read, and what each says about the counterparty behind a rate card.
- Who actually owns the GPUs you are renting — The special-purpose vehicles, leases and guarantees that sit between your invoice and the hardware, and where your prepayment lands in that stack.
- What a three-year-old GPU is worth and who will guarantee it — Depreciation schedules, the thin used-GPU market, and the vendor guarantee that carves out its own silicon.
- Why refinancing is the real uptime risk on a multi-year GPU contract — Five-year loans written against three-year contracts, the 2026 to 2028 maturity cluster, and how to test a provider’s solvency before signing.
- What happens to your workloads and data when a GPU provider fails — Step-in, assignment and change-of-control clauses, and the operational sequence from a missed payment to a dead inference endpoint.
- How much of a premium is a better-capitalised GPU provider worth — Neocloud versus hyperscaler, reserved versus on-demand, and the arithmetic behind paying more for a stronger balance sheet.
Why does a GPU provider’s balance sheet show up in your procurement decision?
The moment you sign a multi-year reservation you stop being a customer and become an unsecured creditor of the provider for the full term. Your workloads sit above the provider’s credit agreement, cross-defaulted, so a default on one facility can trigger the others. This is underwriting a counterparty, and the most accessible external monitor you have is the provider’s credit spread.
A backlog is promised revenue, not cash: contracted revenue before the cost of power, buildings and staff. Diligence answers whether a vendor performs; a reservation asks whether it will still exist.
A spread near 1,000 basis points is treated as the distress line, which CoreWeave’s own CDS has now crossed; AI lending is a quarter of recent corporate credit growth, and the Reserve Bank of Australia flags the same pattern.
Read more: What a 15.78% bond yield reveals about your GPU provider
Who actually owns the GPUs you are renting?
The entity named on your contract is often not the entity that holds the chips or the debt. Special-purpose vehicles sit between your invoice and the hardware, holding title and borrowing. If the answers to three questions, who signs, who owns, who owes, are different, your counterparty is a structure, and your prepayment sits at the bottom of that stack.
The chain runs: equity, senior debt, the vehicle, the hardware, the lease, the operator, the contract, then you. The Apollo and Blackstone vehicle behind Anthropic, roughly $35 billion, bought about a million Google TPUs and leased them to Anthropic, which held neither. Off-balance-sheet commitments take aggregate hyperscaler obligations from roughly $800 billion to $3.5 trillion, so stated debt is a starting point.
Then the prepayment. When an offtaker, a buyer committing to capacity, pays 15 to 25% of contract value at signing, that cash is credit support: an unsecured loan to the operator, repaid in service, absorbing first loss. Protection is a standard ask.
Read more: Who actually owns the GPUs you are renting
What is the hardware actually worth, and how long does the financing last?
Depreciation runs at roughly half of revenue at the largest listed neoclouds, and a six-year schedule on identical silicon lasts 50% longer than a four-year one, a solvency choice rather than an accounting one. Five-year loans are pledged against contracts averaging three years, so refinancing in the 2026 to 2028 window, not hardware failure, is the realistic route to an interruption.
Lenders secure loans against hardware a competitor replaces in eighteen months, and against the customer contracts. CoreWeave books six years, Nebius books four, and the used-GPU market is too thin to settle it: about $26.3 million of sales over three years.
Nvidia’s $105 billion residual value guarantee for OpenAI carves its own equipment out as Guarantor Property. The vendor guarantees everything except the thing that depreciates fastest.
On duration, that gap is refilled by renewal at unknown future prices, inside the 2026 to 2028 cluster and among a small, correlated lender group. The loan outlives the hardware; the collateral question is who pays.
Read more: What a three-year-old GPU is worth and who will guarantee it, and why refinancing is the real uptime risk
What actually happens to your workloads and data if the provider fails?
On enforcement, the lender’s objective is to preserve the cash flows pledged against the loan, and your contract is one of them. Step-in, assignment and change-of-control clauses decide who keeps serving you; egress capacity and data gravity decide whether you can leave. “Investment-grade offtaker” language protects the lender’s repayment schedule, not your continuity, so you need your own contractual protections.
The sequence: missed refinancing or covenant breach, cross-default across facilities, enforcement or administration, operational standstill, notice to customers, assignment or termination. At every step the buyer is an asset being preserved.
Escrow of GPU inventory decides whether your prepayment bought an asset or a claim; step-in and transition rights decide who takes over. “Investment-grade offtaker” is written for the lender: Lambda’s $926 million term loan was justified by it.
Take-or-pay requires payment whether or not capacity is delivered, while the provider’s delivery obligation is capped by service credits that are themselves an unsecured claim. Reserved capacity stops being schedulable, in-flight runs die, and retrieval depends on egress and whether an administrator bills for it. Portability is decided by egress and architectural coupling, not the exit clause.
Read more: What happens to your workloads and data when a GPU provider fails
How much of a premium is a better-capitalised provider worth?
A premium is justified where the expected loss, probability of failure over the term multiplied by migration and downtime cost, exceeds the rate differential across committed megawatt-years. What the premium buys is not a guarantee of service but a lower probability that your contract becomes a line item in an enforcement. Reserved versus on-demand changes the shape of that exposure, not its existence.
Compare providers on four axes: leverage against contracted cash flow, customer concentration, contract duration, and whether the contracting entity carries the debt. Oracle keeps the ranking honest: a BBB rating and $638 billion of remaining performance obligations, yet a 282 basis point CDS, force majeure on Project Jupiter and $18 billion of project debt.
A reservation is credit and price exposure for the whole term; on-demand renews price risk hourly and parks credit risk with the provider. Revenue per monetisable megawatt is the better read: powered-shell landlords monetise roughly $1.2 to $2.3 million, full-stack neoclouds $7 to $13 million, carrying the financing and depreciation risk.
The five levers, term length, commitment size, prepayment protection, portability and provider choice, decide the outcome before any discount.
Read more: How much of a premium is a better-capitalised GPU provider worth
Route by state:
- A rate card in front of you goes to the signal article.
- An unnamed entity on the contract goes to ownership.
- A five-year term goes to collateral and duration.
- An imminent signature goes to enforcement.
- An open choice goes to premium.
Mid-negotiation defaults to duration.
Frequently Asked Questions
What is a neocloud, and how is it different from a hyperscaler?
A neocloud is a specialist GPU operator: a company whose product is accelerated compute and whose fleet is financed specifically to sell it. It does not carry a diversified cloud business or a balance sheet built on decades of software cash flow. That narrowness is the point. A hyperscaler can absorb a bad quarter against other revenue lines, while a neocloud services its debt from GPU rentals alone, which is why its credit spread moves faster than its service quality does.
What is a credit default swap, in plain English?
A credit default swap is insurance on a company’s debt, bought and sold by investors who never lend it a dollar. The buyer pays an annual premium, quoted in basis points, and collects if the borrower defaults. That premium is the spread. When CoreWeave’s five-year CDS sits near 1,038 basis points, the market is charging roughly 10% a year to insure its debt, which tells you what professional investors think of the next five years.
How do I check the credit of a GPU provider that has no bond and no credit rating?
Ask for documents instead, because a private operator still has lenders who have already done the work. Request the debt maturity profile, operating lease obligations, largest customer concentrations, covenant headroom and a summary of the facility agreement. If a provider will not show you its maturity schedule, treat that refusal as the answer. Unrated borrowers typically demand 25 to 30% prepayment precisely because nobody has priced them yet.
Does a blue-chip customer list make a GPU provider safer?
Not by itself. Concentration is a warning as much as a reference: three unnamed customers produced 72% of CoreWeave’s quarterly revenue, and Microsoft walking away from Nscale’s Loughton site weeks before its listing proved that an anchor tenant is a promise rather than a floor. A blue-chip name only becomes credit support when it backstops the debt directly, as Google’s lease guarantee did in the Anthropic vehicle.
What should I ask for in a portability clause?
Ask for export rights measured in days and paid for by the provider, not a general right to leave. Specifically: your model weights and checkpoints in an open format, egress capacity committed at a stated rate during a transition, container images you can rebuild elsewhere, and no exit fee during a credit event. Portability is decided by egress and architecture, so a clause that promises graceful exit while leaving 40 terabytes stranded is decoration.
What does take-or-pay mean, and can I stop paying if the capacity stops working?
Take-or-pay means you owe the commitment whether or not capacity is delivered, so a provider outage does not suspend your invoice. Your remedy is usually limited to service credits, which are themselves an unsecured claim if the provider becomes insolvent. That asymmetry is the core of the credit exposure. If you want payment tied to delivery, the suspension, set-off and termination-for-credit-event language has to be negotiated before signing.
How much prepayment is normal, and how do I protect it?
Expect a request for 15 to 25% of contract value at signing, and 25 to 30% from an unrated borrower, because that cash is credit support rather than a discount. It is an unsecured loan to the operator that absorbs first loss on the facility. Protect it with an amortisation schedule that draws the balance down as service is delivered, escrow or a security interest where offered, and a refund on termination.
Can I negotiate the right to terminate if the provider is downgraded?
Yes, and it is one of the few protections that costs the provider nothing today. Ask for a credit-event termination right triggered by a downgrade below an agreed rating, a CDS spread sustained above a stated level, a covenant breach or a missed payment, with a refund of unearned prepayment. Lenders already build cross-default triggers into their own facilities, so the precedent exists in the capital stack above you.
If the provider’s parent company guarantees the contract, does that solve it?
Only if the guarantor carries the cash flow to honour it. A parent guarantee moves the exposure to another balance sheet, and in leveraged structures the parent is often the thinner entity. Check the guarantor’s own debt, its rating and whether the guarantee is capped or conditional. The Nvidia guaranty on the SB Energy campus is the cautionary version: $105 billion of cover that expressly carves its own equipment out.
Should I spread inference across two providers to reduce credit risk?
Splitting workloads reduces concentration but doubles the integration and egress work, and it does not remove the exposure if both providers borrow from the same lenders and refinance in the same window. Correlated maturities mean correlated refinancing conditions. A second provider earns its keep when the two have different funding sources and different customer bases. Otherwise you have paid for redundancy without buying independence.
Do I need trade credit insurance for a multi-year GPU commitment?
Usually not, and rarely as a substitute for contract terms. Trade credit cover is written for goods delivered on short payment terms, not for a three-year capacity commitment where your loss is migration cost rather than an unpaid invoice. The cheaper protection is structural: shorter term, smaller commitment, protected prepayment and real portability. If you do buy cover, check whether it responds to an insolvency during a service interruption.
